Differences from OpenAI

Nodea runs open models on your own devices, so some behavior differs from OpenAI’s hosted API. If you’re moving code from OpenAI, check this list.

Scope

• Only Chat Completions and Models. There’s no Responses API, legacy Completions, embeddings, images, audio, files, batches or fine-tuning.
• Text only. Image, audio and file inputs and audio output aren’t supported.
• The models are the ones you enable. OpenAI model names like gpt-4o don’t exist. Use a model name from your Models page.

Output

• stop keeps the stop text. OpenAI cuts the output just before the stop sequence; Nodea includes it at the end. Strip it yourself if needed.
• seed is ignored. Use temperature: 0 for repeatable output.
• usage can be missing. Token counts come from the device. If it doesn’t report them, usage is left out rather than estimated. With n > 1, completion_tokens counts one choice, not all of them.
• refusal is always null, and there’s no content moderation, so finish_reason is never "content_filter".

Tool calling

• tool_choice supports "auto" and "none". "required" and naming a specific function return 400.
• parallel_tool_calls: false isn’t supported. The model may return several tool calls at once.
• With tool_choice: "none", some models still write out a tool call as plain text in content.
• Alongside a correct tool call, content may contain leftover formatting from the model’s chat template. Rely on tool_calls, not content, when finish_reason is "tool_calls".
• In streaming, tool calls arrive complete in one chunk instead of incrementally.
• How well tools work depends on the model. Smaller models are less reliable.

Structured output

• With response_format set to json_schema, the output is constrained to the schema during generation, so it always matches.
• Generation can continue past the end of the JSON until it hits the token limit, giving finish_reason: "length" even though the JSON is complete. Set a reasonable max_completion_tokens.

Other parameters

• logit_bias token ids refer to the model’s own tokenizer. Bias maps built for OpenAI’s tokenizer won’t do what you expect.
• reasoning_effort accepts only "low", "medium" and "high".
• Fields Nodea ignores are listed in the x-nodea-ignored-params response header. See Ignored.

Capacity and limits

• One response at a time per device. To handle more requests at once, add devices. See How devices get models.
• Requests can wait for a device. If no device is serving the model, a request waits up to 120 seconds before failing with 503 model_unavailable.
• No rate limits, so no 429 responses or x-ratelimit-* headers.
• Completions aren’t stored. store and metadata are ignored, and there’s no API to retrieve past completions.